Edge deployment imposes strict latency, memory, and energy constraints that scale directly with network depth, yet the question of which architectural paradigm offers the best accuracy-efficiency tradeoff at ultra-low depth (four to six layers) remains open. We present the first systematic comparison of convolutional, pure transformer, and hybrid CNN-transformer models under fixed shallow depth budgets. We develop an analytical framework characterizing the representational cost of local convolutional versus global attention-based feature mixing as a function of depth, and validate it with experiments on ImageNet-1K across architectures spanning MobileNetV2, DeiT, Swin, MobileViT, and EfficientFormer. Beyond FLOPs and accuracy, we report real-world latency and energy measurements on an NVIDIA Jetson Nano and examine deployment feasibility on Cortex-M class microcontrollers. Our experiments reveal how accuracy, latency, and memory scale across all three architecture families as depth decreases, providing practitioners with direct guidance on which model class to choose for a given depth and hardware budget.

This paper proposes a novel type of random finite set (RFS), namely augmented RFS, to address the problem of resolvable group target tracking, which integrates the information of both the group attributes and the dynamic state of group targets into random finite sets. Specifically, we initially introduce an augmented random finite set framework, incorporating group labels and group cardinality to estimate both the trajectories and states of group targets. Then, a new multi-target filter based on the augmented RFS is proposed to achieve the process of group target tracking. Finally, simulation experiments are conducted to demonstrate the effectiveness of the proposed filter in group target tracking scenarios.

Vision Transformers (ViTs) have emerged as a powerful architecture for computer vision tasks due to their ability to model long-range dependencies and global contextual relationships. However, their substantial compute and memory demands hinder efficient deployment in scenarios with strict energy and bandwidth limitations. In this work, we propose Opto-ViT, the first near-sensor, region-aware ViT accelerator leveraging silicon photonics (SiPh) for real-time and energy-efficient vision processing. Opto-ViT features a hybrid electronic-photonic architecture, where the optical core handles compute-intensive matrix multiplications using Vertical-Cavity Surface-Emitting Lasers (VCSELs) and Microring Resonators (MRs), while nonlinear functions and normalization are executed electronically. To reduce redundant computation and patch processing, we introduce a lightweight Mask Generation Network (MGNet) that identifies regions of interest in the current frame and prunes irrelevant patches before ViT encoding. We further co-optimize the ViT backbone using quantization-aware training and matrix decomposition tailored for photonic constraints. Experiments across device fabrication, circuit and architecture co-design, to classification, detection, and video tasks demonstrate that Opto-ViT achieves 100.4 KFPS/W with up to 84% energy savings with less than 1.6% accuracy loss, while enabling scalable and efficient ViT deployment at the edge.

Resolvable group target tracking is of great challenge due to the complex motion interaction between group targets, which leads to tracking performance degradation. To solve this problem, a cardinality-balanced multi-target multi-Bernoulli filter based on the graph theory and leader-follower model is proposed. In the proposed filter, firstly, the group targets are divided into leaders and followers by mean of the leader-follower model. Furthermore, the graph theory is used to establish the state transition equations between those divided group targets. Lastly, the process of state prediction is given, and its corresponding implementation is derived by Gaussian mixture approximations. Simulation experiments verify the superiority and effectiveness of the proposed filter.

The interference-plus-noise covariance matrix (INCM) reconstruction-based adaptive beamforming methods have been successful in preventing signal self-nulling. However, their computational complexity is generally high, which cannot be neglected. In this paper, we propose a data-driven adaptive beamforming method named Deep-Reconstruction, which utilizes deep learning to establish a direct mapping from the sample covariance matrix to the inverse of the INCM. Specifically, we devise a Unet-based fully convolutional network to extract the low-dimensional representations of interferences and noise from the sample covariance matrix. Meanwhile, a conjugate symmetrization layer is designed to maintain a Hermitian structure of the network output. As a result, an accurate estimation of the inverse of the INCM can be obtained for the beamformer design. Simulation results demonstrate that the proposed method can effectively avoid signal self-nulling, while achieving a higher computational efficiency as compared to the traditional methods.

In the integrated sensing and communication system, sensing functionalities are expected to benefit the communication instead of compromising its performance. In this paper, a sensing-aided communication channel estimation method is proposed, where the non-cooperative moving targets are localized and the associated propagation paths are excluded from the channel. Specifically, the received signal is formulated as a high-order tensor and then decomposed for channel parameter estimation. The parameters including velocity and angles of each path are automatically paired in the decomposed tensor factors, which enables identification of the high-velocity paths of moving targets. Then, by excluding the parameters of moving targets, a stable communication channel can be constructed. According to simulation, the proposed method contributes to enhanced data transmission performance while accurately localizing the moving targets.

Compared with the commonly used Cramér-Rao bound, the Ziv-Zakai bound (ZZB) is a global tight lower bound for evaluating the performance of parameter estimators. However, the existing ZZB for multi-source direction-of-arrival (DOA) estimation is derived under an ideal array assumption, where the gain-phase error is not taken into account. Hence, when the gain-phase error exists, the existing ZZB cannot provide a global effective bound. To address this problem, we incorporate the gain-phase error term into the ZZB derivation, and formulate the ZZB as an explicit function of the gain-phase error, which reveals that the gain-phase error affects the ZZB by introducing an extra signal-to-noise ratio gain/loss to the received signals. Simulation results demonstrate that the derived ZZB is global effective and tighter than the existing ZZB for multi-source DOA estimation in gain-phase error scenarios.

The abuse of drones has raised critical concerns about public security and personal privacy, bringing an urgent requirement for drone recognition. Existing radio frequency (RF)-based recognition methods follow the assumption of the closed set, resulting in the unknown signals being misclassified as known classes. To address this problem, we propose a Signal Semantic-based open Set Recognition (S3R) method in this paper. First, the short-time Fourier transform is introduced to construct the signal spectra, decoupling the drone signals with other interference signals. Then, we design a texture extractor and a position extractor to extract the texture features and position features from the spectra, respectively. The extracted features are further fused and structurally optimized to construct distinguishable signal semantics. Based on the structural characteristics of signal semantics, an outlier analysis-based semantic classifier is proposed, which searches the outliers of each known class in the closed set as the bounding thresholds to detect unknown instances. Finally, the detected unknown instances are further classified into their exact classes by implementing clustering in a new semantic space, where semantics are augmented by introducing basic features from the intermediate layers of the texture extractor. Besides, a real-world spectrogram dataset of commonly-used drones is released, which includes 24 classes and covers 7 brands. Extensive experiments demonstrate that the proposed S3R method outperforms the state-of-the-art methods in terms of accuracy and generalizability for both the closed set and the open set.

The rapid development of unmanned aerial vehicle (UAV) market presents potential threats to public security and personal privacy, and the vision sensors are widely deployed to detect the invasive UAVs because of the intuitivity and accessibility of the video. However, the small pixel area and weak morphological characteristics of distant invasive UAVs pose a considerable challenge to the detection precision. Prior work on UAV detection simply focuses on the information fusion between different feature layers, but ignoring the feature information inside each layer. In addition, to detect small UAV in video streams, the motion information of the target is also a noteworthy feature. In this regard, we propose a feature super-resolution-based UAV detector with motion information extractor. The proposed network fully utilizes the motion information of UAVs between temporal frames and the spatial invariant features between different resolution frames to pursue a high-accuracy small UAV detection performance. Experiments on Drone vs Birds dataset are carried out, and it is demonstrated that a higher detection accuracy on small UAVs is achieved compared with the baseline.

Existing matrix-based neural network for direction-of-arrival (DOA) estimation has to train a large amount of parameters proportional to the length of vectorized signal statistics, resulting in a heavy system overload. To address the problem, a tensorized neural layer decomposition-based neural network is proposed for 2-D DOA estimation. In particular, the covariance tensor of tensor signals is propagated to hidden state tensors. The feedforward propagation is formulated as an inverse Tucker decomposition, such that parameters in the tensorized neural layers are compressed into inverse Tucker factors. Accordingly, the tensorized backpropagation procedure is designed for network training. It is proved that the number of parameters is significantly reduced, which leads to a faster training process. Simulation results demonstrate that the proposed method reduces the number of trained parameters by more than 122,000 times compared to the matrix-based neural network while maintaining a moderate accuracy.

The previous robust adaptive beamformer for sparse array enjoys performance improvement due to enhanced degrees-of-freedom (DOFs), while the polarization diversity of signals is not considered. The polarization diversity is an important factor that could be exploited to design a more functional beamformer. Specifically, in this paper, a cascaded sparse array (CSA) composed of diversely polarized antennas, which is polarization sensitive and has reduced mutual coupling, is proposed with a closed-form expression for the array geometry. Then, a polarimetric sparse reconstruction beamformer operating in the joint spatial and polarization domain is proposed for the CSA, and it is capable of suppressing the interferences using the information in the additional polarization domain with enhanced DOFs. As a result, it not only offers reduced costs but also improved functionality, demonstrating its potential value in future wireless communications. The polynomial rooting based joint direction-of-arrival and polarization estimation procedures are then given to support the power distribution estimation in a closed-form manner. Subsequently, the estimated steering vector of the desired signal and the reconstructed interference-plus-noise covariance matrix are combined to calculate the proposed beamformer. Numerical simulations are included to verify the potential advantages of the proposed CSA as well as the superior performance and robustness of the proposed beamformer.

Direction-of-arrival (DOA) estimation using sub-Nyquist tensor signals benefits from enhanced performance by extracting structural angular information with multi-dimensional sparse arrays. Although convolutional neural network (CNN) has been employed to achieve efficient DOA estimation in challenging conditions, conventional methods demand excessive memory storage and computation power to process sub-Nyquist tensor statistics. In this letter, we propose a decomposed CNN for sub-Nyquist tensor-based 2-D DOA estimation, where an augmented coarray tensor is derived and used as the network input. To compress convolution kernels for efficient coarray tensor propagation, we develop a convolution kernel decomposition approach. This enables the acquisition of canonical polyadic (CP) factors containing compressed parameters. Performing decomposable convolution between the coarray tensor and the CP factors leads to resource-efficient DOA estimation. Our simulation results indicate that the proposed method conserves system resources while maintaining competitive performance.

In this paper, a coarray tensor completion-based direction-of-arrival (DOA) estimation method is proposed for coprime planar array. To perform Nyquist-matched coarray signal processing, the completion of the coarray tensor corresponding to an augmented discontinuous virtual array is pursued. However, it is difficult to impose a low-rank regularization on the incomplete coarray tensor with slices of missing elements for its completion. To solve this problem, a structural tensorization approach is designed to reshape the incomplete coarray tensor into one with distributed missing elements. As such, a coarray tensor completion problem based on tensor nuclear norm minimization is formulated to complete these missing elements. By exploiting the filled virtual array obtained from the completed coarray tensor, a super-resolution DOA estimation can be achieved in closed-form.

Conventional tensor direction-of-arrival (DOA) estimation methods for sparse arrays apply canonical polyadic decomposition (CPD) to the high-order coarray covariance tensor for retrieving angle information. However, due to the low convergence rate of CPD-based algorithms for high-order tensors, these methods suffer from a high computation cost. To address this issue, a sub-Nyquist tensor train decomposition (SubTTD)-based DOA estimation method is proposed for a three-dimensional (3-D) sparse array, where an augmented virtual array is derived from the sub-Nyquist tensor statistics. To reduce computational complexity of processing the 6-D coarray covariance tensor, the proposed SubTTD model efficiently decomposes it into a train of head matrix, 3-D core tensors, and tail matrix. Based on that, a core tensor decomposition and a change-of-basis transformation for the head matrix are designed to retrieve canonical polyadic factors of the coarray covariance tensor for DOA estimation. The computational efficiency of the proposed method is theoretically analyzed, and its effectiveness is verified via simulations.

Existing tensor-based coherent direction-of-arrival (DOA) estimation methods adopting spatial smoothing to decorrelate the coherent tensor statistics usually lead to a poor decorrelation performance. In this letter, we propose a structured tensor reconstruction method for two-dimensional coherent DOA estimation, which then avoids the inefficient spatial smoothing. In particular, after investigating the structural property of the four-dimensional incoherent covariance tensor, we propose a tensorial Hermitian Toeplitz mapping rule to reconstruct a structured covariance tensor from the rank-deficient coherent covariance tensor statistics. It is theoretically proved that, the reconstructed covariance tensor admits a decorrelated canonical polyadic model with a tensorial Hermitian Toeplitz structure, whose decomposition ensures a closed-form coherent DOA estimation. The effectiveness of the proposed method is verified by simulations.

Conventional canonical polyadic decomposition (CPD) approach for tensor-based sparse array direction-of-arrival (DOA) estimation typically partitions the coarray statistics to generate a full-rank coarray tensor for decomposition. However, such an operation ignores the spatial relevance among the partitioned coarray statistics. In this letter, we propose a coupled coarray tensor CPD-based two-dimensional DOA estimation method for a specially designed coprime L-shaped array. In particular, a shifting coarray concatenation approach is developed to factorize the partitioned fourth-order coarray statistics into multiple coupled coarray tensors. To make full use of the inherent spatial relevance among these coarray tensors, a coupled coarray tensor CPD approach is proposed to jointly decompose them for high-accuracy DOA estimation in a closed-form manner. According to the uniqueness condition analysis on the coupled coarray tensor CPD, an increased number of degrees-of-freedom for the proposed method is guaranteed.

Coprime arrays can cope with the underdetermined case for direction-of-arrival (DOA) estimation. However, the popular matrix-based coarray signal processing approaches suffer performance loss on the underlying characteristics among the multi-dimensional signals. To address this problem, we propose a novel coarray tensor-based two-dimensional underdetermined DOA estimation method for coprime planar array in this paper, where both the multi-dimensional information of the received signals and the augmented coarray are effectively utilized. The received signal tensors of the coprime planar array are constructed by concatenating each snapshot, which are then transformed to an augmented uniform rectangular array statistics for extending the effective array aperture. Subsequently, the tensorization technique is adopted to increase the degrees-of-freedom of the proposed DOA estimation method, and a structured coarray tensor is optimized for super-resolution DOA estimation. The effectiveness of the proposed method is verified via simulation results.

The frequency diverse array (FDA) radar has been extensively studied due to its unique range‐angle dependent beampattern. Time‐invariant spatial patterns have been reported for FDAs proposed so far. However, some recent studies indicate that the patterns obtained by using these methodologies neglect the time‐range or frequency‐phase relationship, and the definition of time in some equations are misinterpreted which results in erroneous conclusions. Taking the time‐variant property of FDA beampatterns into consideration, in this study, the authors propose a short range FDA radar to generate a time‐variant focused range‐angle dependent transmit beampattern, where chirp waveform with multi‐carrier FDA architecture is used. Both the fixed and time‐modulated frequency offsets are considered to analyse the proposed FDA radar. The frequency offset employed across each element is generated by chaos sequence and sine function. By compensating the propagation delays of signals, the transmitted signals sum up constructively to focus on the desired range‐angle sector only at a specific instant of time. Numerical results are implemented to verify the validity of the proposed schemes.

Coprime arrays are capable of achieving an increased number of degrees-of-freedom by operating the coarray signals. However, their non-uniform coarrays prevent the full utilization of the available signals. To address this problem, a novel coarray interpolation-based direction-of-arrival (DOA) estimation algorithm via covariance matrix reconstruction is proposed in this paper. In particular, we formulate a gridless optimization problem to reconstruct the covariance matrix of the interpolated coarray, such that all the coarray observations are fully utilized. We also investigate the rotational invariance in the coarray domain to retrieve the DOAs. Neither spatial sampling nor spectrum searching is required in the proposed algorithm, indicating the capability of resolving off-grid DOAs. Simulation results demonstrate the effectiveness of the proposed DOA estimation algorithm.

In this letter, we propose a coprime array interpolation approach to provide an off-grid direction-of-arrival (DOA) estimation. Through array interpolation, a uniform linear array (ULA) with the same aperture is generated from the deterministic non-uniform coprime array. Taking the observed correlations calculated from the signals received at the coprime array, a gridless convex optimization problem is formulated to recover all the rows and columns of the unknown correlation matrix entries corresponding to the interpolated sensors. The optimized Hermitian positive semidefinite Toeplitz matrix functions as the covariance matrix of the interpolated ULA, which enables to resolve off-grid sources. Simulation results demonstrate that the proposed array interpolation-based DOA estimation algorithm achieves improved performance as compared to existing coarray-based DOA estimation algorithms in terms of the number of achievable degrees-of-freedom and estimation accuracy.

A coprime array enables an increased number of degrees-of-freedom by deriving a non-uniform virtual array. However, existing work such as spatial smoothing fails to utilize all of the information provided by the coprime array, which results in performance loss. In this paper, we propose a novel coprime virtual array interpolation-based direction- of-arrival (DOA) estimation algorithm by Toeplitz matrix reconstruction. After investigating the challenges caused by the non-uniformity, we introduce the idea of array interpolation to construct a uniform linear virtual array, such that the information contained in the coprime virtual array can be fully utilized. According to the statistics of non-uniform coprime virtual array signal, we formulate a convex optimization problem for DOA estimation by reconstructing the covariance matrix of the equivalent received signals of the interpolated coprime virtual array under the Toeplitz constraint. Simulation results demonstrate the effectiveness of the proposed algorithm.

A coprime array has a larger array aperture as well as increased degrees‐of‐freedom (DOFs), compared with a uniform linear array with the same number of physical sensors. Therefore, in a practical wireless communication system, it is capable to provide desirable performance with a low‐computational complexity. In this study, the authors focus on the problem of efficient direction‐of‐arrival (DOA) estimation, where a coprime array is incorporated with the idea of compressive sensing. Specifically, the authors first generate a random compressive sensing kernel to compress the received signals of coprime array to lower‐dimensional measurements, which can be viewed as a sketch of the original received signals. The compressed measurements are subsequently utilised to perform high‐resolution DOA estimation, where the large array aperture of the coprime array is maintained. Moreover, the authors also utilise the derived equivalent virtual array signal of the compressed measurements for DOA estimation, where the superiority of coprime array in achieving a higher number of DOFs can be retained. Theoretical analyses and simulation results verify the effectiveness of the proposed methods in terms of computational complexity, resolution, and the number of DOFs.

In this paper, we propose a novel adaptive beamforming algorithm for coprime array by compressive sensing the virtual uniform linear array signal. Based on the idea of coprime sampling, a much longer virtual uniform linear array can be generated from a coprime array. With a compressive sensing matrix, a connection can be built between the coprime array with fewer physical sensors and the virtual uniform linear array with much more virtual sensors. Hence, the proposed adaptive beamforming algorithm takes full advantage of the longer virtual array. The performance increment provided by the virtual array is much larger than the performance loss due to the introduced compressive sensing. Hence, the beam-former using the virtual array is expected to obtain much better performance than those using the coprime array directly. Simulation results demonstrate the effectiveness of the proposed adaptive beamforming algorithm.

In this paper, we propose a novel robust adaptive beamforming algorithm with direction-of-arrival (DOA) support for the coprime array. Specifically, by using the property of coprime number, we may estimate the DOAs of sources by matching two super-resolution spatial spectra of the pair of decomposed coprime subarrays. After that, the power of each source can be estimated via a covariance matrix joint estimation problem corresponding to the pair of decomposed coprime sub-arrays. Taking the estimated DOAs and their corresponding power as the support information, the interference-plus-noise covariance matrix for the coprime array can be reconstructed, from which the minimum variance distortionless response beamformer weight vector can be calculated. Simulation results show that the proposed adaptive beamforming algorithm is more robust to signal look direction mismatch than the existing algorithms.

In this paper, we propose a direction-of-arrival estimation method by covariance matrix sparse reconstruction of coprime array. Specifically, source locations are estimated by solving a newly formulated convex optimization problem, where the difference between the spatially smoothed covariance matrix and the sparsely reconstructed one is minimized. Then, a sliding window scheme is designed for source enumeration. Finally, the power of each source is re-estimated as a least squares problem. Compared with existing methods, the proposed method achieves more accurate source localization and power estimation performance with full utilization of increased degrees of freedom provided by coprime array.